Clothes Segmentation in Consumer Photos by Using Information

• Main Authors: Chian-Ta Hong, 洪建達
• Other Authors: Wei-Ta Chu
• Format: Others
• Language: zh-TW
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/70512160799085419822

碩士 === 國立中正大學 === 資訊工程所 === 97 === At present, image segmentation is still a difficult research problem. The reasons are that regions of image suffer from object occlusion, region complexity, and the impact of variation of brightness. This thesis develops robust image segmentation techniques to extract clothing regions from images. Due to clothing locations are different in images, it is hard to extract features for them. Therefore, we propose a method to establish clothing mask automatically. At the same time, we exploit the same person wears the same clothing in a short period, and then establish common information between images. Then, we exploit clothing mask to calculate clothing features and non-clothing features for each image. We use the technology of graph cuts to extract clothing regions accurately. In experiments, we effectively overcome the problems of pose, object occlusion, complexity of clothing and variation of brightness to achieve clothing segmentation. In the future, we will propose better features and apply this module in other related research, such as object segmentation or retrieval.